Dependencies between morphophonological patterns
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
Dependencies between morphophonological patterns
Guy Tabachnick (New York University), guyt@nyu.edu
April 30, 2021
1 Introduction
One problem that learners face: how to arbitrarily associate morphophonological patterns with specific
lexical items. Examples:
• English past tense (Albright and Hayes, 2003): reach–reached but teach–taught
• Russian diminutives (Gouskova et al., 2015): [ángj il]–[ángj il-ók] ‘angel’, but [mónstr]—[mónstr-j ik]
‘monster’
When learners need to generate a previously unseen form (past tense, diminutive) of a novel lexical item,
how do they do it?
• Typical approaches focus on phonological generalizations that speakers can pick up on
– -ok never attaches to stems ending in consonant clusters
• Ackerman et al. (2009); Ackerman and Malouf (2013): inflectional systems are organized in a way to
allow speakers to efficiently predict one form given others (solving the Paradigm Cell Filling Problem)
Some models of lexically specific patterns, like the sublexicon model (Gouskova et al., 2015; Becker and
Gouskova, 2016), allow speakers to pick up on phonological biases.
Today: a case study from Hungarian showing that just looking at the phonology isn’t enough:
• nouns take one of two possessive suffixes, -6 or -j6
• nouns in a class known as “lowering stems” overwhelmingly take -6 and only rarely -j6
Crucially: these correlations are not due to shared phonological characteristics between the two classes!
Models that capture phonological tendencies in a pattern are not capturing the whole story.
I propose an extension to the sublexicon model to account for these morphological depencies.
1PhoNE Guy Tabachnick
Dependencies between morphophonological patterns April 30, 2021
2 Case study: Hungarian possessives
2.1 Possessive allomorphy
In Hungarian, the possessive suffix has two allomorphs (ignoring vowel harmony), -6/E and -j6/jE. Nouns
select one or the other, with some allowing both (sometimes with a distinction in meaning):
word gloss possessive
pa:r ‘pair’ pa:r-j6
sor ‘line’ Sor-6
6bl6k ‘window’ 6bl6k-(j)6
Table 1: Some Hungarian possessive forms (see Rácz and Rebrus, 2012)
Rácz and Rebrus (2012) summarize: -j6 has spread to become the productive form today (used for loan-
words, etc.)
Some basic generalizations about possessive allomorph selection (they describe other non-categorical pat-
terns as well):
stem shape possessive allomorphy example
V-final always -j6 k6pu-j6 ‘gate’
palatal-final always -6 la:ñ-6 ‘girl’
sibilant-final always -6 ga:z-6 ‘gas’
Table 2: Rácz and Rebrus (2012): Generalizations about possessive allomorph selection
2.2 Lowering stems
They also note: irregular noun classes usually take -6. One example: lowering stems, which have a low
vowel in forms like the plural where most nouns have a mid vowel (see also Rebrus, 2013; Rebrus et al.,
2017):1
stem class example gloss plural possessive
regular pa:r ‘pair’ pa:r-ok pa:r-j6
lowering éa:r ‘factory’ éa:r-6k éa:r-6
Table 3: Hungarian lowering stems
1 Bydefault, nouns are non-lowering stems, while all but a handful of adjectives are lowering stems, and some forms behave
differently depending on the syntactic environment (Rebrus and Szigetvári, 2018). Setting aside these issues, I limit my inquiry to
words that are listed as nouns.
2PhoNE Guy Tabachnick
Dependencies between morphophonological patterns April 30, 2021
How strong is this tendency? Quite!2
possessive possessive
stem class
-6/E -j6/jE -6/E -j6/jE
regular 642 935 ka:r-ok, ka:r-6 ‘damage’ pa:r-ok, pa:r-j6 ‘pair’
lowering 104 15 éa:r-6k, éa:r-6 ‘factory’ ña:r-6k, ña:r-j6 ‘poplar’
Table 4: Distribution of Hungarian lowering stems and possessive allomorphy
One possibility: lowering stems just look like typical -6 words (maybe a lot end in sibilants and palatals, for
example).
This may be a small factor, for example:
• 470/1575 ≈ 30% of regular stems end in palatals or sibilants (which only take -6)
• 48/119 ≈ 40% of lowering stems do
However, this alone doesn’t explain the effect—the morphological effect is independent of the phonological
ones!
2.3 Phonological and morphological factors influencing possessive allomorphy
I fit two models for predicting a noun’s possessive allomorph—one with only phonological factors, one with
stem class as a predictor as well.3
Not only does adding stem class improve the model, but the phonological effect sizes stay steady—the
morphological dependency is independent!
Example: the effect of final consonant place. (Here higher coefficient → more likely to take -j6 relative to
the baseline, alveolar. All effects are significant, p < .0001.)
effect size in model effect size in model
predictor predicted rates
without stem class with stem class
final C: alveolar (baseline)
final C: labial −2.05 −2.25 tsi:m-jE < si:n-jE
final C: palatal −9.01 −9.40 ryé-jE < tSyd-jE
final C: velar −3.35 −3.63 gø:g-jE < tSø:d-jE
stem class: regular (baseline)
stem class: lowering — −3.77 ña:r-j6 < pa:r-j6
Table 5: The effect of final consonant place on Hungarian possessive allomorphy, with and without stem
class as a predictor
2 MyHungarian source is Papp (1969), a morphological dictionary. I consider consonant-final, monomorphemic nouns that are
not marked as variable in the possessive. Not included in Table 4 are 23 nouns listed as having variable stem class.
3 The models were fit using the lrm function in R’s rms package (R Core Team, 2020; Harrell Jr., 2020). See Appendix A for
the full models.
3PhoNE Guy Tabachnick
Dependencies between morphophonological patterns April 30, 2021
Not only is being a lowering stem a fairly strong predictor of selecting -6, but it gives information that the
phonology of a noun alone does not!
To put it another way: seeing a noun’s phonological form is helpful to the learner in figuring out its
possessive—but seeing its plural to learn that it’s a lowering stem is even more helpful.
3 A sublexicon model with morphological dependencies
3.1 The basic sublexicon model
The phonological effects in Table 5 can be handled with a sublexicon model (Gouskova et al., 2015; Becker
and Gouskova, 2016).
Steps for the learner (in the basic sublexicon model):
1. The learner associates a morpheme with a particular morphosyntactic context and creates a rule of
exponence linking form to meaning
• basic rule of exponence for the Hungarian possessive: POSS ↔ j6
2. The learner encounters the same morpheme with a different exponent, and splits the rule of exponence
into two contextually specific rules marked by diacritics4
• spun-off rules of exponence for the Hungarian possessive:
– POSS ↔ j6 / [+j] ___
– POSS ↔ 6 / [−j] ___
3. The learner places lexical items into sublexicons associated with these diacritics (i.e. their behavior in
the given context)
• [+j] sublexicon = {pa:r, k6pu, . . .}
• [−j] sublexicon = {éa:r, ga:z, . . .}
4. The learner generates a phonotactic grammar (Hayes and Wilson, 2008) for each sublexicon, which
captures generalizations over the phonological shape of its members through phonotactic constraints
• highly weighted constraints in the [+j] sublexicon: *[palatal]#, *[sibilant]#
• highly weighted constraints in the [−j] sublexicon: *V#
5. When the learner encounters a new lexeme and has to generate the form in question, they evaluate the
stem against each sublexicon’s grammar to see which is a better fit and place it accordingly
• nonce word [o:d] rates better on the [+j] sublexicon
• the learner marks it with [+j], places it in the [+j] sublexicon, and produces [o:d-j6]5
4 For the sake of illustration, I mark both new rules of exponence with a diacritic. An alternative is that one rule gets the diacritic
and the other is left as a default applying in other cases.
5 By a regular rule of palatal assimilation, the actual surface form would be [o:é6].
4PhoNE Guy Tabachnick
Dependencies between morphophonological patterns April 30, 2021
3.2 Adding morphological dependencies
Meanwhile, the learner has also noticed that there’s variation in the vowel in plural and other forms and has
placed words into additional sublexicons based on a [±lower] diacritic:6
• PL ↔ 6k / [+lower] ___ [+lower] = {[éa:r], [ña:r], . . .}
• PL ↔ ok / [−lower] ___ [−lower] = {[ka:r], [pa:r], . . .}
Recall: being a lowering stem is a very good predictor for taking -6 as a possessive allomorph ([éa:r-6k] →
[éa:r-6]).
In sublexicon terms: if a noun is in the [+lower] sublexicon, it’s likely to also be in the [−j] sublexicon.
The basic sublexicon model, however, can’t account for this! Let’s add a step:
40 . When creating a new sublexicon, add *S UBLEX constraints penalizing membership in already ex-
isting sublexicons, and add *S UBLEX constraints for the new sublexicon into all existing sublexical
grammars.
– the new [±j] sublexicons are seeded with *S UBLEX-[+lower] and *S UBLEX-[−lower]
– by definition: *S UBLEX-[+lower] penalizes lexical items in the [+lower] sublexicon
– *S UBLEX-[+lower] will be weighted heavily in the [+j] sublexicon grammar: candidates for
taking possessive -j6 are penalized for also being a lowering stem with plural -6k
The addition of *S UBLEX constraints thus extends the sublexicon model to pick up not just on phonological
patterns, but also on dependencies between lexically specific patterns!
3.3 The organization of morphological systems
Sublexicons can help us conceptualize learning declension classes, which are essentially diacritics that
appear in multiple contextual rules of exponence:
• in Russian, class II nouns like [komnata] ‘room’ have [-a] in the nominative, [-u] in the accusative,
etc. (Corbett and Fraser, 1993)
• potential learning process:
– learner starts off with [NOM:a] and [ACC:u], etc. diacritics/sublexicons
– the two classes contain almost identical nouns and grammars
– accordingly, the learner merges them into a single [class:II] sublexicon
6 Theserules of exponence are simplified for the purposes of illustration. A given noun’s choice of vowel, known as a “linking
vowel”, is uniform across several morphological categories, of which plural is only one, so there is good reason to think that the
selection of vowel is handled in a different rule of exponence from insertion of the plural allomorph.
5PhoNE Guy Tabachnick
Dependencies between morphophonological patterns April 30, 2021
Relatedly: researchers like Ackerman et al. (2009); Ackerman and Malouf (2013) have argued for treating
the inflectional paradigm as a basic unit of linguistic structure based on the ways in which inflected forms
are informative about one another.
• some patterns that seem almost unlearnably complex from the perspective of “X contains an exponent
of morpheme M” look much more efficient when viewed as “X is a member of paradigm P”
Sublexicons with *S UBLEX constraints allow the formal grammatical model to capture some of this infor-
mativity, even under theories of exponence like Distributed Morphology that deny the formal existence of
paradigms.
4 Summary
Hungarian possessives show:
• many patterns of lexically specified behavior show phonological biases that speakers can pick up on
and exploit in generating new forms
• sometimes two such patterns are informative about one another—no reason speakers can’t exploit
these as well
• we can model these dependencies using the sublexicon approach with additional *S UBLEX con-
straints, which give us a way to pick up on biases in the organization of a language’s morphological
system
6PhoNE Guy Tabachnick
Dependencies between morphophonological patterns April 30, 2021
A Appendix: Full models
This appendix contains the full models calculated for predicting possessive allomorphs among monomor-
phemic Hungarian nouns without variable possessive forms, with data taken from Papp (1969). The models
were calculated using the lrm function in R’s rms package (R Core Team, 2020; Harrell Jr., 2020). The first
model (Table 6) contains only phonological factors, while the second (Table 7) also includes the morpho-
logical factor of lowering stem class.
I assembled the models using stepwise comparison, where each additional factor significantly improved the
model (p < .0001) according to the same package’s lrtest function. For both cases, an ANOVA confirmed
that each factor in the final models was significant (p < .0001). Furthermore, each addition to the models
improved their Akaike Information Criterion (AIC), a metric that rewards model likelihood but penalizes
model complexity—that is, the improvement in model fit brought by adding each factor outweighed the cost
of having a more complex model.
β coef SE Wald z p predicted rates
Intercept 3.03 .32 9.48 bot-6
C Manner (default: plosive)
fricative −1.40 .39 −3.54 .0004 ra:f-j6 < la:p-j6
sibilant −11.57 1.09 −10.65PhoNE Guy Tabachnick
Dependencies between morphophonological patterns April 30, 2021
β coef SE Wald z p predicted rates
Intercept 3.54 .34 10.51 bot-6
Stem class (default: non-lowering)
lowering −3.77 .45 −8.47PhoNE Guy Tabachnick
Dependencies between morphophonological patterns April 30, 2021
References
Ackerman, F., Blevins, J. P., and Malouf, R. (2009). Parts and wholes: Implicative patterns in inflectional
paradigms. In Blevins, J. P. and Blevins, J., editors, Analogy in Grammar: Form and Acquisition, chap-
ter 3, pages 54–82. Oxford University Press, Oxford.
Ackerman, F. and Malouf, R. (2013). Morphological organization: The low conditional entropy conjecture.
Language, 89(3):429–464.
Albright, A. and Hayes, B. (2003). Rules vs. analogy in English past tenses: a computational/experimental
study. Cognition, 90(2):119–161.
Becker, M. and Gouskova, M. (2016). Source-oriented generalizations as grammar inference in Russian
vowel deletion. Linguistic Inquiry, 47(3):391–425.
Corbett, G. G. and Fraser, N. M. (1993). Network Morphology: a DATR account of Russian nominal
inflection. Journal of Linguistics, 29(1):113–142.
Gouskova, M., Newlin-Łukowicz, L., and Kasyanenko, S. (2015). Selectional restrictions as phonotactics
over sublexicons. Lingua, 167:41–81.
Harrell Jr., F. E. (2020). rms: Regression Modeling Strategies. R package version 6.1-0.
Hayes, B. and Wilson, C. (2008). A maximum entropy model of phonotactics and phonotactic learning.
Linguistic Inquiry, 39(3):379–440.
Papp, F. (1969). A magyar nyelv szóvégmutató szótára. Akadémiai Kiadó, Budapest.
R Core Team (2020). R: A Language and Environment for Statistical Computing. R Foundation for Statis-
tical Computing, Vienna.
Rácz, P. and Rebrus, P. (2012). Variation in the possessive allomorphy of Hungarian. In Kiefer, F., Ladányi,
M., and Siptár, P., editors, Current Issues in Morphological Theory: (Ir)regularity, analogy and fre-
quency: Selected papers from the 14th International Morphology Meeting, Budapest, 13–16 May 2010,
volume 322 of Current Issues in Linguistic Theory, pages 51–64, Amsterdam/Philadelphia. John Ben-
jamins.
Rebrus, P. (2013). Miért nincs j? In Benő, A., Fazakas, E., and Kádár, E., editors, „. . .hogy legyen a víznek
lefolyása. . .”: Köszöntő kötet Szilágyi N. Sándor tiszteletére, pages 383–401. Erdélyi Múzeum-Egyesület,
Kolozsvár [Cluj-Napoca].
Rebrus, P. and Szigetvári, P. (2018). Morphosyntactically conditioned morphophonological variation. Pre-
sentation at PhonVar 1, Barcelona.
Rebrus, P., Szigetvári, P., and Törkenczy, M. (2017). Asymmetric variation. In Lindsey, G. and Nevins, A.,
editors, Sonic Signatures: Studies dedicated to John Harris, volume 14 of Language Faculty and Beyond,
pages 164–187. John Benjamins, Amsterdam / Philadelphia.
9You can also read
